Technical Guide | Financial Services

    Shadow AI at Bank Fintech and BaaS Partners: Extending Oversight Past Your Perimeter

    Shadow AI at fintech and BaaS partners is best treated as an evidence and enforcement gap, not a discovery or questionnaire problem. Contracts, due diligence and periodic reviews describe partner AI use at a point in time. Runtime policy enforcement in the execution path can show what AI was allowed to do and what it did for the interactions and tool calls it covers. A sound approach uses both and documents what neither reaches.

    Evaluation criteria for audit-ready partner AI evidence

    • Does the control act on individual interactions and tool calls in the execution path, or only review activity afterward?
    • Is each decision recorded with the policy applied, the action taken (allow, redact, block, rate limit, escalate) and the identity of any human approver?
    • Is the coverage boundary documented, including AI traffic that bypasses the control point?
    • Can the control be self-managed in the cloud environment the data owner chooses, with configurable retention and data residency?
    • Does the evidence persist in a form the bank's third-party risk and model risk teams can query without relying on the partner to produce it?
    • Are claims limited to interaction-level and tool-call-level enforcement, without implying model lineage or full workflow tracing?

    Where oversight breaks down past the perimeter

    The accountability position is settled. The interagency guidance on third-party relationships, issued by the FDIC, FRB and OCC on June 6, 2023, states that a banking organization's use of third parties does not diminish or remove its responsibility to operate safely and soundly and to comply with applicable laws, including those on consumer protection and security of customer information. As of November 2026, a proposal issued September 11, 2026 by the FDIC, OCC, FRB and NCUA would replace that guidance with a version tailored more closely to the level of risk. It keeps the same principle on responsibility. Banks should track the proposal, but the direction on accountability is consistent.

    The guidance is risk-based and technology-neutral. It does not set out AI-specific requirements for how a sponsor bank watches a partner's AI tools. The OCC has said AI risk management should be consistent with conventional model risk management practices. It has also noted that services under the Bank Service Company Act could, depending on facts and circumstances, include banking services supported by AI. One OCC supervisory framework document also lists interactions with fintech or agentic systems the bank does not control as a potential channel risk.

    Taken together, these statements put partner AI inside the bank's responsibility. The harder question is operational: how does the bank know what AI a partner is actually using when the partner operates the environment?

    How shadow AI arises in partner relationships

    Shadow AI in this context is AI use that a partner's own approval process, or the bank's third-party risk program, has not accounted for. It rarely looks like a deliberate decision. It accumulates through ordinary work. The common pathways are general, not specific to banking.

    Public AI sites are the simplest: an employee pastes a support transcript, a dispute file or a spreadsheet into a consumer chat tool. Coding assistants can send source code, configuration and sometimes credentials or sample data to an external model. Agents and MCP servers add tool access, so an agent can read from or write to internal systems under its own sequence of calls. Unsanctioned model APIs, called directly from partner applications, can move customer records or underwriting inputs to a model nobody reviewed.

    What data can flow depends on the partner's architecture. In a BaaS arrangement that can include bank customer data, customer interactions and inputs to compliance-relevant processes such as onboarding or complaint handling. A questionnaire answered once a year cannot track tools that are adopted and replaced in weeks. The gap is not the partner's honesty. It is that point-in-time descriptions and day-to-day behavior are different things.

    What contracts and due diligence provide versus runtime enforcement

    The OCC has described expectations of robust due diligence, effective contract management and ongoing oversight, scaled to the criticality of the service. The first two are pre-conditions. The open question is what ongoing oversight can show about day-to-day AI use. The comparison below separates the layers so neither is credited with what it cannot deliver.

    • Timing: Contracts and diligence: set at onboarding and revisited periodically. Runtime enforcement: applied to each interaction or tool call as it happens.
    • What it shows: Contracts and diligence: stated policy, intended AI use and attestations. Runtime enforcement: what was allowed, redacted, blocked, rate limited or escalated.
    • Response to a violation: Contracts and diligence: remedies and findings after the fact. Runtime enforcement: policy enforced at runtime, not reviewed after the fact.
    • Coverage: Contracts and diligence: whatever the partner discloses. Runtime enforcement: only traffic routed through the control point.
    • Human decisions: Contracts and diligence: approval processes described on paper. Runtime enforcement: escalation for human approval recorded per action.
    • Evidence type: Contracts and diligence: documents and reports. Runtime enforcement: continuous records generated by enforcement itself.

    What runtime enforcement can and cannot cover

    Enforcement in the execution path means a policy decision is made on each interaction or tool call before it completes. The available actions are allow, redact, block, rate limit and escalate for human approval. A prompt containing customer identifiers might be redacted. A call from an agent to a system outside an approved set might be blocked. A high-impact action might be held for a named approver. The scope is individual interactions and tool calls. This is not transaction-level governance, and it does not trace an agentic workflow end to end.

    The boundary matters as much as the capability. A control point sees only the traffic routed through it. AI use that bypasses it, such as a personal device on a personal network, is outside its coverage. A bank should expect that limit to be written down and tied to the partner's network and access architecture, so an examiner is not left to infer coverage.

    Cross-organizational deployment is the other practical constraint. The partner operates the environment, so the bank needs clarity on who deploys the control, who can change a policy, and where evidence is stored. Self-managed deployment in the customer cloud, with configurable retention and data residency, lets the entity that owns the data decide where evidence sits and how long it is kept. These are architectural questions to settle in the partner agreement, not assumptions.

    Connecting evidence to third-party and model risk programs

    Scope partner oversight by criticality, in line with the risk-based direction of both the 2023 guidance and the 2026 proposal. A partner that handles customer interactions or compliance-relevant processes warrants closer runtime evidence than one using AI for internal drafting. Partner AI that touches bank data or customers should then be mapped into existing model risk management practice, consistent with the OCC's position that AI follows conventional model risk management. Runtime evidence shows how a model or tool was used and what controls applied. It does not replace model validation, and it does not provide model lineage.

    For banks weighing proof, note that Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, and produces audit-ready evidence mapped to frameworks including NYDFS and NIST. It has run in a regulated enterprise in continuous production for more than 18 months, with 99.95% uptime on live production traffic. Insurance, a regulated financial sector, is where Trussed is deepest. Evaluate any vendor, including Trussed, against the criteria above, and confirm with your regulators and counsel how the evidence fits your own program.

    Two layers of partner AI oversight

    Contracts and due diligence

    Point-in-time description of intended partner AI use.

    Runtime enforcement

    Allow, redact, block, rate limit or escalate individual interactions and tool calls.

    Evidence

    A record of what partner AI was allowed to do and what it did.

    Documented boundary

    Clear statement of AI use the control point does not see.

    Review your partner AI evidence gap

    Talk through where partner AI use falls outside your current oversight and what runtime evidence would need to show.

    Talk to an Expert
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.